xref: /openbmc/linux/arch/x86/events/intel/lbr.c (revision 715f23b6)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/perf_event.h>
3 #include <linux/types.h>
4 
5 #include <asm/perf_event.h>
6 #include <asm/msr.h>
7 #include <asm/insn.h>
8 
9 #include "../perf_event.h"
10 
11 enum {
12 	LBR_FORMAT_32		= 0x00,
13 	LBR_FORMAT_LIP		= 0x01,
14 	LBR_FORMAT_EIP		= 0x02,
15 	LBR_FORMAT_EIP_FLAGS	= 0x03,
16 	LBR_FORMAT_EIP_FLAGS2	= 0x04,
17 	LBR_FORMAT_INFO		= 0x05,
18 	LBR_FORMAT_TIME		= 0x06,
19 	LBR_FORMAT_MAX_KNOWN    = LBR_FORMAT_TIME,
20 };
21 
22 static const enum {
23 	LBR_EIP_FLAGS		= 1,
24 	LBR_TSX			= 2,
25 } lbr_desc[LBR_FORMAT_MAX_KNOWN + 1] = {
26 	[LBR_FORMAT_EIP_FLAGS]  = LBR_EIP_FLAGS,
27 	[LBR_FORMAT_EIP_FLAGS2] = LBR_EIP_FLAGS | LBR_TSX,
28 };
29 
30 /*
31  * Intel LBR_SELECT bits
32  * Intel Vol3a, April 2011, Section 16.7 Table 16-10
33  *
34  * Hardware branch filter (not available on all CPUs)
35  */
36 #define LBR_KERNEL_BIT		0 /* do not capture at ring0 */
37 #define LBR_USER_BIT		1 /* do not capture at ring > 0 */
38 #define LBR_JCC_BIT		2 /* do not capture conditional branches */
39 #define LBR_REL_CALL_BIT	3 /* do not capture relative calls */
40 #define LBR_IND_CALL_BIT	4 /* do not capture indirect calls */
41 #define LBR_RETURN_BIT		5 /* do not capture near returns */
42 #define LBR_IND_JMP_BIT		6 /* do not capture indirect jumps */
43 #define LBR_REL_JMP_BIT		7 /* do not capture relative jumps */
44 #define LBR_FAR_BIT		8 /* do not capture far branches */
45 #define LBR_CALL_STACK_BIT	9 /* enable call stack */
46 
47 /*
48  * Following bit only exists in Linux; we mask it out before writing it to
49  * the actual MSR. But it helps the constraint perf code to understand
50  * that this is a separate configuration.
51  */
52 #define LBR_NO_INFO_BIT	       63 /* don't read LBR_INFO. */
53 
54 #define LBR_KERNEL	(1 << LBR_KERNEL_BIT)
55 #define LBR_USER	(1 << LBR_USER_BIT)
56 #define LBR_JCC		(1 << LBR_JCC_BIT)
57 #define LBR_REL_CALL	(1 << LBR_REL_CALL_BIT)
58 #define LBR_IND_CALL	(1 << LBR_IND_CALL_BIT)
59 #define LBR_RETURN	(1 << LBR_RETURN_BIT)
60 #define LBR_REL_JMP	(1 << LBR_REL_JMP_BIT)
61 #define LBR_IND_JMP	(1 << LBR_IND_JMP_BIT)
62 #define LBR_FAR		(1 << LBR_FAR_BIT)
63 #define LBR_CALL_STACK	(1 << LBR_CALL_STACK_BIT)
64 #define LBR_NO_INFO	(1ULL << LBR_NO_INFO_BIT)
65 
66 #define LBR_PLM (LBR_KERNEL | LBR_USER)
67 
68 #define LBR_SEL_MASK	0x3ff	/* valid bits in LBR_SELECT */
69 #define LBR_NOT_SUPP	-1	/* LBR filter not supported */
70 #define LBR_IGN		0	/* ignored */
71 
72 #define LBR_ANY		 \
73 	(LBR_JCC	|\
74 	 LBR_REL_CALL	|\
75 	 LBR_IND_CALL	|\
76 	 LBR_RETURN	|\
77 	 LBR_REL_JMP	|\
78 	 LBR_IND_JMP	|\
79 	 LBR_FAR)
80 
81 #define LBR_FROM_FLAG_MISPRED	BIT_ULL(63)
82 #define LBR_FROM_FLAG_IN_TX	BIT_ULL(62)
83 #define LBR_FROM_FLAG_ABORT	BIT_ULL(61)
84 
85 #define LBR_FROM_SIGNEXT_2MSB	(BIT_ULL(60) | BIT_ULL(59))
86 
87 /*
88  * x86control flow change classification
89  * x86control flow changes include branches, interrupts, traps, faults
90  */
91 enum {
92 	X86_BR_NONE		= 0,      /* unknown */
93 
94 	X86_BR_USER		= 1 << 0, /* branch target is user */
95 	X86_BR_KERNEL		= 1 << 1, /* branch target is kernel */
96 
97 	X86_BR_CALL		= 1 << 2, /* call */
98 	X86_BR_RET		= 1 << 3, /* return */
99 	X86_BR_SYSCALL		= 1 << 4, /* syscall */
100 	X86_BR_SYSRET		= 1 << 5, /* syscall return */
101 	X86_BR_INT		= 1 << 6, /* sw interrupt */
102 	X86_BR_IRET		= 1 << 7, /* return from interrupt */
103 	X86_BR_JCC		= 1 << 8, /* conditional */
104 	X86_BR_JMP		= 1 << 9, /* jump */
105 	X86_BR_IRQ		= 1 << 10,/* hw interrupt or trap or fault */
106 	X86_BR_IND_CALL		= 1 << 11,/* indirect calls */
107 	X86_BR_ABORT		= 1 << 12,/* transaction abort */
108 	X86_BR_IN_TX		= 1 << 13,/* in transaction */
109 	X86_BR_NO_TX		= 1 << 14,/* not in transaction */
110 	X86_BR_ZERO_CALL	= 1 << 15,/* zero length call */
111 	X86_BR_CALL_STACK	= 1 << 16,/* call stack */
112 	X86_BR_IND_JMP		= 1 << 17,/* indirect jump */
113 
114 	X86_BR_TYPE_SAVE	= 1 << 18,/* indicate to save branch type */
115 
116 };
117 
118 #define X86_BR_PLM (X86_BR_USER | X86_BR_KERNEL)
119 #define X86_BR_ANYTX (X86_BR_NO_TX | X86_BR_IN_TX)
120 
121 #define X86_BR_ANY       \
122 	(X86_BR_CALL    |\
123 	 X86_BR_RET     |\
124 	 X86_BR_SYSCALL |\
125 	 X86_BR_SYSRET  |\
126 	 X86_BR_INT     |\
127 	 X86_BR_IRET    |\
128 	 X86_BR_JCC     |\
129 	 X86_BR_JMP	 |\
130 	 X86_BR_IRQ	 |\
131 	 X86_BR_ABORT	 |\
132 	 X86_BR_IND_CALL |\
133 	 X86_BR_IND_JMP  |\
134 	 X86_BR_ZERO_CALL)
135 
136 #define X86_BR_ALL (X86_BR_PLM | X86_BR_ANY)
137 
138 #define X86_BR_ANY_CALL		 \
139 	(X86_BR_CALL		|\
140 	 X86_BR_IND_CALL	|\
141 	 X86_BR_ZERO_CALL	|\
142 	 X86_BR_SYSCALL		|\
143 	 X86_BR_IRQ		|\
144 	 X86_BR_INT)
145 
146 static void intel_pmu_lbr_filter(struct cpu_hw_events *cpuc);
147 
148 /*
149  * We only support LBR implementations that have FREEZE_LBRS_ON_PMI
150  * otherwise it becomes near impossible to get a reliable stack.
151  */
152 
153 static void __intel_pmu_lbr_enable(bool pmi)
154 {
155 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
156 	u64 debugctl, lbr_select = 0, orig_debugctl;
157 
158 	/*
159 	 * No need to unfreeze manually, as v4 can do that as part
160 	 * of the GLOBAL_STATUS ack.
161 	 */
162 	if (pmi && x86_pmu.version >= 4)
163 		return;
164 
165 	/*
166 	 * No need to reprogram LBR_SELECT in a PMI, as it
167 	 * did not change.
168 	 */
169 	if (cpuc->lbr_sel)
170 		lbr_select = cpuc->lbr_sel->config & x86_pmu.lbr_sel_mask;
171 	if (!pmi && cpuc->lbr_sel)
172 		wrmsrl(MSR_LBR_SELECT, lbr_select);
173 
174 	rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
175 	orig_debugctl = debugctl;
176 	debugctl |= DEBUGCTLMSR_LBR;
177 	/*
178 	 * LBR callstack does not work well with FREEZE_LBRS_ON_PMI.
179 	 * If FREEZE_LBRS_ON_PMI is set, PMI near call/return instructions
180 	 * may cause superfluous increase/decrease of LBR_TOS.
181 	 */
182 	if (!(lbr_select & LBR_CALL_STACK))
183 		debugctl |= DEBUGCTLMSR_FREEZE_LBRS_ON_PMI;
184 	if (orig_debugctl != debugctl)
185 		wrmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
186 }
187 
188 static void __intel_pmu_lbr_disable(void)
189 {
190 	u64 debugctl;
191 
192 	rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
193 	debugctl &= ~(DEBUGCTLMSR_LBR | DEBUGCTLMSR_FREEZE_LBRS_ON_PMI);
194 	wrmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
195 }
196 
197 static void intel_pmu_lbr_reset_32(void)
198 {
199 	int i;
200 
201 	for (i = 0; i < x86_pmu.lbr_nr; i++)
202 		wrmsrl(x86_pmu.lbr_from + i, 0);
203 }
204 
205 static void intel_pmu_lbr_reset_64(void)
206 {
207 	int i;
208 
209 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
210 		wrmsrl(x86_pmu.lbr_from + i, 0);
211 		wrmsrl(x86_pmu.lbr_to   + i, 0);
212 		if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO)
213 			wrmsrl(MSR_LBR_INFO_0 + i, 0);
214 	}
215 }
216 
217 void intel_pmu_lbr_reset(void)
218 {
219 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
220 
221 	if (!x86_pmu.lbr_nr)
222 		return;
223 
224 	if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_32)
225 		intel_pmu_lbr_reset_32();
226 	else
227 		intel_pmu_lbr_reset_64();
228 
229 	cpuc->last_task_ctx = NULL;
230 	cpuc->last_log_id = 0;
231 }
232 
233 /*
234  * TOS = most recently recorded branch
235  */
236 static inline u64 intel_pmu_lbr_tos(void)
237 {
238 	u64 tos;
239 
240 	rdmsrl(x86_pmu.lbr_tos, tos);
241 	return tos;
242 }
243 
244 enum {
245 	LBR_NONE,
246 	LBR_VALID,
247 };
248 
249 /*
250  * For formats with LBR_TSX flags (e.g. LBR_FORMAT_EIP_FLAGS2), bits 61:62 in
251  * MSR_LAST_BRANCH_FROM_x are the TSX flags when TSX is supported, but when
252  * TSX is not supported they have no consistent behavior:
253  *
254  *   - For wrmsr(), bits 61:62 are considered part of the sign extension.
255  *   - For HW updates (branch captures) bits 61:62 are always OFF and are not
256  *     part of the sign extension.
257  *
258  * Therefore, if:
259  *
260  *   1) LBR has TSX format
261  *   2) CPU has no TSX support enabled
262  *
263  * ... then any value passed to wrmsr() must be sign extended to 63 bits and any
264  * value from rdmsr() must be converted to have a 61 bits sign extension,
265  * ignoring the TSX flags.
266  */
267 static inline bool lbr_from_signext_quirk_needed(void)
268 {
269 	int lbr_format = x86_pmu.intel_cap.lbr_format;
270 	bool tsx_support = boot_cpu_has(X86_FEATURE_HLE) ||
271 			   boot_cpu_has(X86_FEATURE_RTM);
272 
273 	return !tsx_support && (lbr_desc[lbr_format] & LBR_TSX);
274 }
275 
276 static DEFINE_STATIC_KEY_FALSE(lbr_from_quirk_key);
277 
278 /* If quirk is enabled, ensure sign extension is 63 bits: */
279 inline u64 lbr_from_signext_quirk_wr(u64 val)
280 {
281 	if (static_branch_unlikely(&lbr_from_quirk_key)) {
282 		/*
283 		 * Sign extend into bits 61:62 while preserving bit 63.
284 		 *
285 		 * Quirk is enabled when TSX is disabled. Therefore TSX bits
286 		 * in val are always OFF and must be changed to be sign
287 		 * extension bits. Since bits 59:60 are guaranteed to be
288 		 * part of the sign extension bits, we can just copy them
289 		 * to 61:62.
290 		 */
291 		val |= (LBR_FROM_SIGNEXT_2MSB & val) << 2;
292 	}
293 	return val;
294 }
295 
296 /*
297  * If quirk is needed, ensure sign extension is 61 bits:
298  */
299 static u64 lbr_from_signext_quirk_rd(u64 val)
300 {
301 	if (static_branch_unlikely(&lbr_from_quirk_key)) {
302 		/*
303 		 * Quirk is on when TSX is not enabled. Therefore TSX
304 		 * flags must be read as OFF.
305 		 */
306 		val &= ~(LBR_FROM_FLAG_IN_TX | LBR_FROM_FLAG_ABORT);
307 	}
308 	return val;
309 }
310 
311 static inline void wrlbr_from(unsigned int idx, u64 val)
312 {
313 	val = lbr_from_signext_quirk_wr(val);
314 	wrmsrl(x86_pmu.lbr_from + idx, val);
315 }
316 
317 static inline void wrlbr_to(unsigned int idx, u64 val)
318 {
319 	wrmsrl(x86_pmu.lbr_to + idx, val);
320 }
321 
322 static inline u64 rdlbr_from(unsigned int idx)
323 {
324 	u64 val;
325 
326 	rdmsrl(x86_pmu.lbr_from + idx, val);
327 
328 	return lbr_from_signext_quirk_rd(val);
329 }
330 
331 static inline u64 rdlbr_to(unsigned int idx)
332 {
333 	u64 val;
334 
335 	rdmsrl(x86_pmu.lbr_to + idx, val);
336 
337 	return val;
338 }
339 
340 static void __intel_pmu_lbr_restore(struct x86_perf_task_context *task_ctx)
341 {
342 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
343 	int i;
344 	unsigned lbr_idx, mask;
345 	u64 tos;
346 
347 	if (task_ctx->lbr_callstack_users == 0 ||
348 	    task_ctx->lbr_stack_state == LBR_NONE) {
349 		intel_pmu_lbr_reset();
350 		return;
351 	}
352 
353 	tos = task_ctx->tos;
354 	/*
355 	 * Does not restore the LBR registers, if
356 	 * - No one else touched them, and
357 	 * - Did not enter C6
358 	 */
359 	if ((task_ctx == cpuc->last_task_ctx) &&
360 	    (task_ctx->log_id == cpuc->last_log_id) &&
361 	    rdlbr_from(tos)) {
362 		task_ctx->lbr_stack_state = LBR_NONE;
363 		return;
364 	}
365 
366 	mask = x86_pmu.lbr_nr - 1;
367 	for (i = 0; i < task_ctx->valid_lbrs; i++) {
368 		lbr_idx = (tos - i) & mask;
369 		wrlbr_from(lbr_idx, task_ctx->lbr_from[i]);
370 		wrlbr_to  (lbr_idx, task_ctx->lbr_to[i]);
371 
372 		if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO)
373 			wrmsrl(MSR_LBR_INFO_0 + lbr_idx, task_ctx->lbr_info[i]);
374 	}
375 
376 	for (; i < x86_pmu.lbr_nr; i++) {
377 		lbr_idx = (tos - i) & mask;
378 		wrlbr_from(lbr_idx, 0);
379 		wrlbr_to(lbr_idx, 0);
380 		if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO)
381 			wrmsrl(MSR_LBR_INFO_0 + lbr_idx, 0);
382 	}
383 
384 	wrmsrl(x86_pmu.lbr_tos, tos);
385 	task_ctx->lbr_stack_state = LBR_NONE;
386 }
387 
388 static void __intel_pmu_lbr_save(struct x86_perf_task_context *task_ctx)
389 {
390 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
391 	unsigned lbr_idx, mask;
392 	u64 tos, from;
393 	int i;
394 
395 	if (task_ctx->lbr_callstack_users == 0) {
396 		task_ctx->lbr_stack_state = LBR_NONE;
397 		return;
398 	}
399 
400 	mask = x86_pmu.lbr_nr - 1;
401 	tos = intel_pmu_lbr_tos();
402 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
403 		lbr_idx = (tos - i) & mask;
404 		from = rdlbr_from(lbr_idx);
405 		if (!from)
406 			break;
407 		task_ctx->lbr_from[i] = from;
408 		task_ctx->lbr_to[i]   = rdlbr_to(lbr_idx);
409 		if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO)
410 			rdmsrl(MSR_LBR_INFO_0 + lbr_idx, task_ctx->lbr_info[i]);
411 	}
412 	task_ctx->valid_lbrs = i;
413 	task_ctx->tos = tos;
414 	task_ctx->lbr_stack_state = LBR_VALID;
415 
416 	cpuc->last_task_ctx = task_ctx;
417 	cpuc->last_log_id = ++task_ctx->log_id;
418 }
419 
420 void intel_pmu_lbr_swap_task_ctx(struct perf_event_context *prev,
421 				 struct perf_event_context *next)
422 {
423 	struct x86_perf_task_context *prev_ctx_data, *next_ctx_data;
424 
425 	swap(prev->task_ctx_data, next->task_ctx_data);
426 
427 	/*
428 	 * Architecture specific synchronization makes sense in
429 	 * case both prev->task_ctx_data and next->task_ctx_data
430 	 * pointers are allocated.
431 	 */
432 
433 	prev_ctx_data = next->task_ctx_data;
434 	next_ctx_data = prev->task_ctx_data;
435 
436 	if (!prev_ctx_data || !next_ctx_data)
437 		return;
438 
439 	swap(prev_ctx_data->lbr_callstack_users,
440 	     next_ctx_data->lbr_callstack_users);
441 }
442 
443 void intel_pmu_lbr_sched_task(struct perf_event_context *ctx, bool sched_in)
444 {
445 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
446 	struct x86_perf_task_context *task_ctx;
447 
448 	if (!cpuc->lbr_users)
449 		return;
450 
451 	/*
452 	 * If LBR callstack feature is enabled and the stack was saved when
453 	 * the task was scheduled out, restore the stack. Otherwise flush
454 	 * the LBR stack.
455 	 */
456 	task_ctx = ctx ? ctx->task_ctx_data : NULL;
457 	if (task_ctx) {
458 		if (sched_in)
459 			__intel_pmu_lbr_restore(task_ctx);
460 		else
461 			__intel_pmu_lbr_save(task_ctx);
462 		return;
463 	}
464 
465 	/*
466 	 * Since a context switch can flip the address space and LBR entries
467 	 * are not tagged with an identifier, we need to wipe the LBR, even for
468 	 * per-cpu events. You simply cannot resolve the branches from the old
469 	 * address space.
470 	 */
471 	if (sched_in)
472 		intel_pmu_lbr_reset();
473 }
474 
475 static inline bool branch_user_callstack(unsigned br_sel)
476 {
477 	return (br_sel & X86_BR_USER) && (br_sel & X86_BR_CALL_STACK);
478 }
479 
480 void intel_pmu_lbr_add(struct perf_event *event)
481 {
482 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
483 	struct x86_perf_task_context *task_ctx;
484 
485 	if (!x86_pmu.lbr_nr)
486 		return;
487 
488 	cpuc->br_sel = event->hw.branch_reg.reg;
489 
490 	if (branch_user_callstack(cpuc->br_sel) && event->ctx->task_ctx_data) {
491 		task_ctx = event->ctx->task_ctx_data;
492 		task_ctx->lbr_callstack_users++;
493 	}
494 
495 	/*
496 	 * Request pmu::sched_task() callback, which will fire inside the
497 	 * regular perf event scheduling, so that call will:
498 	 *
499 	 *  - restore or wipe; when LBR-callstack,
500 	 *  - wipe; otherwise,
501 	 *
502 	 * when this is from __perf_event_task_sched_in().
503 	 *
504 	 * However, if this is from perf_install_in_context(), no such callback
505 	 * will follow and we'll need to reset the LBR here if this is the
506 	 * first LBR event.
507 	 *
508 	 * The problem is, we cannot tell these cases apart... but we can
509 	 * exclude the biggest chunk of cases by looking at
510 	 * event->total_time_running. An event that has accrued runtime cannot
511 	 * be 'new'. Conversely, a new event can get installed through the
512 	 * context switch path for the first time.
513 	 */
514 	if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip > 0)
515 		cpuc->lbr_pebs_users++;
516 	perf_sched_cb_inc(event->ctx->pmu);
517 	if (!cpuc->lbr_users++ && !event->total_time_running)
518 		intel_pmu_lbr_reset();
519 }
520 
521 void intel_pmu_lbr_del(struct perf_event *event)
522 {
523 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
524 	struct x86_perf_task_context *task_ctx;
525 
526 	if (!x86_pmu.lbr_nr)
527 		return;
528 
529 	if (branch_user_callstack(cpuc->br_sel) &&
530 	    event->ctx->task_ctx_data) {
531 		task_ctx = event->ctx->task_ctx_data;
532 		task_ctx->lbr_callstack_users--;
533 	}
534 
535 	if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip > 0)
536 		cpuc->lbr_pebs_users--;
537 	cpuc->lbr_users--;
538 	WARN_ON_ONCE(cpuc->lbr_users < 0);
539 	WARN_ON_ONCE(cpuc->lbr_pebs_users < 0);
540 	perf_sched_cb_dec(event->ctx->pmu);
541 }
542 
543 void intel_pmu_lbr_enable_all(bool pmi)
544 {
545 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
546 
547 	if (cpuc->lbr_users)
548 		__intel_pmu_lbr_enable(pmi);
549 }
550 
551 void intel_pmu_lbr_disable_all(void)
552 {
553 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
554 
555 	if (cpuc->lbr_users)
556 		__intel_pmu_lbr_disable();
557 }
558 
559 static void intel_pmu_lbr_read_32(struct cpu_hw_events *cpuc)
560 {
561 	unsigned long mask = x86_pmu.lbr_nr - 1;
562 	u64 tos = intel_pmu_lbr_tos();
563 	int i;
564 
565 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
566 		unsigned long lbr_idx = (tos - i) & mask;
567 		union {
568 			struct {
569 				u32 from;
570 				u32 to;
571 			};
572 			u64     lbr;
573 		} msr_lastbranch;
574 
575 		rdmsrl(x86_pmu.lbr_from + lbr_idx, msr_lastbranch.lbr);
576 
577 		cpuc->lbr_entries[i].from	= msr_lastbranch.from;
578 		cpuc->lbr_entries[i].to		= msr_lastbranch.to;
579 		cpuc->lbr_entries[i].mispred	= 0;
580 		cpuc->lbr_entries[i].predicted	= 0;
581 		cpuc->lbr_entries[i].in_tx	= 0;
582 		cpuc->lbr_entries[i].abort	= 0;
583 		cpuc->lbr_entries[i].cycles	= 0;
584 		cpuc->lbr_entries[i].type	= 0;
585 		cpuc->lbr_entries[i].reserved	= 0;
586 	}
587 	cpuc->lbr_stack.nr = i;
588 }
589 
590 /*
591  * Due to lack of segmentation in Linux the effective address (offset)
592  * is the same as the linear address, allowing us to merge the LIP and EIP
593  * LBR formats.
594  */
595 static void intel_pmu_lbr_read_64(struct cpu_hw_events *cpuc)
596 {
597 	bool need_info = false, call_stack = false;
598 	unsigned long mask = x86_pmu.lbr_nr - 1;
599 	int lbr_format = x86_pmu.intel_cap.lbr_format;
600 	u64 tos = intel_pmu_lbr_tos();
601 	int i;
602 	int out = 0;
603 	int num = x86_pmu.lbr_nr;
604 
605 	if (cpuc->lbr_sel) {
606 		need_info = !(cpuc->lbr_sel->config & LBR_NO_INFO);
607 		if (cpuc->lbr_sel->config & LBR_CALL_STACK)
608 			call_stack = true;
609 	}
610 
611 	for (i = 0; i < num; i++) {
612 		unsigned long lbr_idx = (tos - i) & mask;
613 		u64 from, to, mis = 0, pred = 0, in_tx = 0, abort = 0;
614 		int skip = 0;
615 		u16 cycles = 0;
616 		int lbr_flags = lbr_desc[lbr_format];
617 
618 		from = rdlbr_from(lbr_idx);
619 		to   = rdlbr_to(lbr_idx);
620 
621 		/*
622 		 * Read LBR call stack entries
623 		 * until invalid entry (0s) is detected.
624 		 */
625 		if (call_stack && !from)
626 			break;
627 
628 		if (lbr_format == LBR_FORMAT_INFO && need_info) {
629 			u64 info;
630 
631 			rdmsrl(MSR_LBR_INFO_0 + lbr_idx, info);
632 			mis = !!(info & LBR_INFO_MISPRED);
633 			pred = !mis;
634 			in_tx = !!(info & LBR_INFO_IN_TX);
635 			abort = !!(info & LBR_INFO_ABORT);
636 			cycles = (info & LBR_INFO_CYCLES);
637 		}
638 
639 		if (lbr_format == LBR_FORMAT_TIME) {
640 			mis = !!(from & LBR_FROM_FLAG_MISPRED);
641 			pred = !mis;
642 			skip = 1;
643 			cycles = ((to >> 48) & LBR_INFO_CYCLES);
644 
645 			to = (u64)((((s64)to) << 16) >> 16);
646 		}
647 
648 		if (lbr_flags & LBR_EIP_FLAGS) {
649 			mis = !!(from & LBR_FROM_FLAG_MISPRED);
650 			pred = !mis;
651 			skip = 1;
652 		}
653 		if (lbr_flags & LBR_TSX) {
654 			in_tx = !!(from & LBR_FROM_FLAG_IN_TX);
655 			abort = !!(from & LBR_FROM_FLAG_ABORT);
656 			skip = 3;
657 		}
658 		from = (u64)((((s64)from) << skip) >> skip);
659 
660 		/*
661 		 * Some CPUs report duplicated abort records,
662 		 * with the second entry not having an abort bit set.
663 		 * Skip them here. This loop runs backwards,
664 		 * so we need to undo the previous record.
665 		 * If the abort just happened outside the window
666 		 * the extra entry cannot be removed.
667 		 */
668 		if (abort && x86_pmu.lbr_double_abort && out > 0)
669 			out--;
670 
671 		cpuc->lbr_entries[out].from	 = from;
672 		cpuc->lbr_entries[out].to	 = to;
673 		cpuc->lbr_entries[out].mispred	 = mis;
674 		cpuc->lbr_entries[out].predicted = pred;
675 		cpuc->lbr_entries[out].in_tx	 = in_tx;
676 		cpuc->lbr_entries[out].abort	 = abort;
677 		cpuc->lbr_entries[out].cycles	 = cycles;
678 		cpuc->lbr_entries[out].type	 = 0;
679 		cpuc->lbr_entries[out].reserved	 = 0;
680 		out++;
681 	}
682 	cpuc->lbr_stack.nr = out;
683 }
684 
685 void intel_pmu_lbr_read(void)
686 {
687 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
688 
689 	/*
690 	 * Don't read when all LBRs users are using adaptive PEBS.
691 	 *
692 	 * This could be smarter and actually check the event,
693 	 * but this simple approach seems to work for now.
694 	 */
695 	if (!cpuc->lbr_users || cpuc->lbr_users == cpuc->lbr_pebs_users)
696 		return;
697 
698 	if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_32)
699 		intel_pmu_lbr_read_32(cpuc);
700 	else
701 		intel_pmu_lbr_read_64(cpuc);
702 
703 	intel_pmu_lbr_filter(cpuc);
704 }
705 
706 /*
707  * SW filter is used:
708  * - in case there is no HW filter
709  * - in case the HW filter has errata or limitations
710  */
711 static int intel_pmu_setup_sw_lbr_filter(struct perf_event *event)
712 {
713 	u64 br_type = event->attr.branch_sample_type;
714 	int mask = 0;
715 
716 	if (br_type & PERF_SAMPLE_BRANCH_USER)
717 		mask |= X86_BR_USER;
718 
719 	if (br_type & PERF_SAMPLE_BRANCH_KERNEL)
720 		mask |= X86_BR_KERNEL;
721 
722 	/* we ignore BRANCH_HV here */
723 
724 	if (br_type & PERF_SAMPLE_BRANCH_ANY)
725 		mask |= X86_BR_ANY;
726 
727 	if (br_type & PERF_SAMPLE_BRANCH_ANY_CALL)
728 		mask |= X86_BR_ANY_CALL;
729 
730 	if (br_type & PERF_SAMPLE_BRANCH_ANY_RETURN)
731 		mask |= X86_BR_RET | X86_BR_IRET | X86_BR_SYSRET;
732 
733 	if (br_type & PERF_SAMPLE_BRANCH_IND_CALL)
734 		mask |= X86_BR_IND_CALL;
735 
736 	if (br_type & PERF_SAMPLE_BRANCH_ABORT_TX)
737 		mask |= X86_BR_ABORT;
738 
739 	if (br_type & PERF_SAMPLE_BRANCH_IN_TX)
740 		mask |= X86_BR_IN_TX;
741 
742 	if (br_type & PERF_SAMPLE_BRANCH_NO_TX)
743 		mask |= X86_BR_NO_TX;
744 
745 	if (br_type & PERF_SAMPLE_BRANCH_COND)
746 		mask |= X86_BR_JCC;
747 
748 	if (br_type & PERF_SAMPLE_BRANCH_CALL_STACK) {
749 		if (!x86_pmu_has_lbr_callstack())
750 			return -EOPNOTSUPP;
751 		if (mask & ~(X86_BR_USER | X86_BR_KERNEL))
752 			return -EINVAL;
753 		mask |= X86_BR_CALL | X86_BR_IND_CALL | X86_BR_RET |
754 			X86_BR_CALL_STACK;
755 	}
756 
757 	if (br_type & PERF_SAMPLE_BRANCH_IND_JUMP)
758 		mask |= X86_BR_IND_JMP;
759 
760 	if (br_type & PERF_SAMPLE_BRANCH_CALL)
761 		mask |= X86_BR_CALL | X86_BR_ZERO_CALL;
762 
763 	if (br_type & PERF_SAMPLE_BRANCH_TYPE_SAVE)
764 		mask |= X86_BR_TYPE_SAVE;
765 
766 	/*
767 	 * stash actual user request into reg, it may
768 	 * be used by fixup code for some CPU
769 	 */
770 	event->hw.branch_reg.reg = mask;
771 	return 0;
772 }
773 
774 /*
775  * setup the HW LBR filter
776  * Used only when available, may not be enough to disambiguate
777  * all branches, may need the help of the SW filter
778  */
779 static int intel_pmu_setup_hw_lbr_filter(struct perf_event *event)
780 {
781 	struct hw_perf_event_extra *reg;
782 	u64 br_type = event->attr.branch_sample_type;
783 	u64 mask = 0, v;
784 	int i;
785 
786 	for (i = 0; i < PERF_SAMPLE_BRANCH_MAX_SHIFT; i++) {
787 		if (!(br_type & (1ULL << i)))
788 			continue;
789 
790 		v = x86_pmu.lbr_sel_map[i];
791 		if (v == LBR_NOT_SUPP)
792 			return -EOPNOTSUPP;
793 
794 		if (v != LBR_IGN)
795 			mask |= v;
796 	}
797 
798 	reg = &event->hw.branch_reg;
799 	reg->idx = EXTRA_REG_LBR;
800 
801 	/*
802 	 * The first 9 bits (LBR_SEL_MASK) in LBR_SELECT operate
803 	 * in suppress mode. So LBR_SELECT should be set to
804 	 * (~mask & LBR_SEL_MASK) | (mask & ~LBR_SEL_MASK)
805 	 * But the 10th bit LBR_CALL_STACK does not operate
806 	 * in suppress mode.
807 	 */
808 	reg->config = mask ^ (x86_pmu.lbr_sel_mask & ~LBR_CALL_STACK);
809 
810 	if ((br_type & PERF_SAMPLE_BRANCH_NO_CYCLES) &&
811 	    (br_type & PERF_SAMPLE_BRANCH_NO_FLAGS) &&
812 	    (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO))
813 		reg->config |= LBR_NO_INFO;
814 
815 	return 0;
816 }
817 
818 int intel_pmu_setup_lbr_filter(struct perf_event *event)
819 {
820 	int ret = 0;
821 
822 	/*
823 	 * no LBR on this PMU
824 	 */
825 	if (!x86_pmu.lbr_nr)
826 		return -EOPNOTSUPP;
827 
828 	/*
829 	 * setup SW LBR filter
830 	 */
831 	ret = intel_pmu_setup_sw_lbr_filter(event);
832 	if (ret)
833 		return ret;
834 
835 	/*
836 	 * setup HW LBR filter, if any
837 	 */
838 	if (x86_pmu.lbr_sel_map)
839 		ret = intel_pmu_setup_hw_lbr_filter(event);
840 
841 	return ret;
842 }
843 
844 /*
845  * return the type of control flow change at address "from"
846  * instruction is not necessarily a branch (in case of interrupt).
847  *
848  * The branch type returned also includes the priv level of the
849  * target of the control flow change (X86_BR_USER, X86_BR_KERNEL).
850  *
851  * If a branch type is unknown OR the instruction cannot be
852  * decoded (e.g., text page not present), then X86_BR_NONE is
853  * returned.
854  */
855 static int branch_type(unsigned long from, unsigned long to, int abort)
856 {
857 	struct insn insn;
858 	void *addr;
859 	int bytes_read, bytes_left;
860 	int ret = X86_BR_NONE;
861 	int ext, to_plm, from_plm;
862 	u8 buf[MAX_INSN_SIZE];
863 	int is64 = 0;
864 
865 	to_plm = kernel_ip(to) ? X86_BR_KERNEL : X86_BR_USER;
866 	from_plm = kernel_ip(from) ? X86_BR_KERNEL : X86_BR_USER;
867 
868 	/*
869 	 * maybe zero if lbr did not fill up after a reset by the time
870 	 * we get a PMU interrupt
871 	 */
872 	if (from == 0 || to == 0)
873 		return X86_BR_NONE;
874 
875 	if (abort)
876 		return X86_BR_ABORT | to_plm;
877 
878 	if (from_plm == X86_BR_USER) {
879 		/*
880 		 * can happen if measuring at the user level only
881 		 * and we interrupt in a kernel thread, e.g., idle.
882 		 */
883 		if (!current->mm)
884 			return X86_BR_NONE;
885 
886 		/* may fail if text not present */
887 		bytes_left = copy_from_user_nmi(buf, (void __user *)from,
888 						MAX_INSN_SIZE);
889 		bytes_read = MAX_INSN_SIZE - bytes_left;
890 		if (!bytes_read)
891 			return X86_BR_NONE;
892 
893 		addr = buf;
894 	} else {
895 		/*
896 		 * The LBR logs any address in the IP, even if the IP just
897 		 * faulted. This means userspace can control the from address.
898 		 * Ensure we don't blindy read any address by validating it is
899 		 * a known text address.
900 		 */
901 		if (kernel_text_address(from)) {
902 			addr = (void *)from;
903 			/*
904 			 * Assume we can get the maximum possible size
905 			 * when grabbing kernel data.  This is not
906 			 * _strictly_ true since we could possibly be
907 			 * executing up next to a memory hole, but
908 			 * it is very unlikely to be a problem.
909 			 */
910 			bytes_read = MAX_INSN_SIZE;
911 		} else {
912 			return X86_BR_NONE;
913 		}
914 	}
915 
916 	/*
917 	 * decoder needs to know the ABI especially
918 	 * on 64-bit systems running 32-bit apps
919 	 */
920 #ifdef CONFIG_X86_64
921 	is64 = kernel_ip((unsigned long)addr) || !test_thread_flag(TIF_IA32);
922 #endif
923 	insn_init(&insn, addr, bytes_read, is64);
924 	insn_get_opcode(&insn);
925 	if (!insn.opcode.got)
926 		return X86_BR_ABORT;
927 
928 	switch (insn.opcode.bytes[0]) {
929 	case 0xf:
930 		switch (insn.opcode.bytes[1]) {
931 		case 0x05: /* syscall */
932 		case 0x34: /* sysenter */
933 			ret = X86_BR_SYSCALL;
934 			break;
935 		case 0x07: /* sysret */
936 		case 0x35: /* sysexit */
937 			ret = X86_BR_SYSRET;
938 			break;
939 		case 0x80 ... 0x8f: /* conditional */
940 			ret = X86_BR_JCC;
941 			break;
942 		default:
943 			ret = X86_BR_NONE;
944 		}
945 		break;
946 	case 0x70 ... 0x7f: /* conditional */
947 		ret = X86_BR_JCC;
948 		break;
949 	case 0xc2: /* near ret */
950 	case 0xc3: /* near ret */
951 	case 0xca: /* far ret */
952 	case 0xcb: /* far ret */
953 		ret = X86_BR_RET;
954 		break;
955 	case 0xcf: /* iret */
956 		ret = X86_BR_IRET;
957 		break;
958 	case 0xcc ... 0xce: /* int */
959 		ret = X86_BR_INT;
960 		break;
961 	case 0xe8: /* call near rel */
962 		insn_get_immediate(&insn);
963 		if (insn.immediate1.value == 0) {
964 			/* zero length call */
965 			ret = X86_BR_ZERO_CALL;
966 			break;
967 		}
968 		/* fall through */
969 	case 0x9a: /* call far absolute */
970 		ret = X86_BR_CALL;
971 		break;
972 	case 0xe0 ... 0xe3: /* loop jmp */
973 		ret = X86_BR_JCC;
974 		break;
975 	case 0xe9 ... 0xeb: /* jmp */
976 		ret = X86_BR_JMP;
977 		break;
978 	case 0xff: /* call near absolute, call far absolute ind */
979 		insn_get_modrm(&insn);
980 		ext = (insn.modrm.bytes[0] >> 3) & 0x7;
981 		switch (ext) {
982 		case 2: /* near ind call */
983 		case 3: /* far ind call */
984 			ret = X86_BR_IND_CALL;
985 			break;
986 		case 4:
987 		case 5:
988 			ret = X86_BR_IND_JMP;
989 			break;
990 		}
991 		break;
992 	default:
993 		ret = X86_BR_NONE;
994 	}
995 	/*
996 	 * interrupts, traps, faults (and thus ring transition) may
997 	 * occur on any instructions. Thus, to classify them correctly,
998 	 * we need to first look at the from and to priv levels. If they
999 	 * are different and to is in the kernel, then it indicates
1000 	 * a ring transition. If the from instruction is not a ring
1001 	 * transition instr (syscall, systenter, int), then it means
1002 	 * it was a irq, trap or fault.
1003 	 *
1004 	 * we have no way of detecting kernel to kernel faults.
1005 	 */
1006 	if (from_plm == X86_BR_USER && to_plm == X86_BR_KERNEL
1007 	    && ret != X86_BR_SYSCALL && ret != X86_BR_INT)
1008 		ret = X86_BR_IRQ;
1009 
1010 	/*
1011 	 * branch priv level determined by target as
1012 	 * is done by HW when LBR_SELECT is implemented
1013 	 */
1014 	if (ret != X86_BR_NONE)
1015 		ret |= to_plm;
1016 
1017 	return ret;
1018 }
1019 
1020 #define X86_BR_TYPE_MAP_MAX	16
1021 
1022 static int branch_map[X86_BR_TYPE_MAP_MAX] = {
1023 	PERF_BR_CALL,		/* X86_BR_CALL */
1024 	PERF_BR_RET,		/* X86_BR_RET */
1025 	PERF_BR_SYSCALL,	/* X86_BR_SYSCALL */
1026 	PERF_BR_SYSRET,		/* X86_BR_SYSRET */
1027 	PERF_BR_UNKNOWN,	/* X86_BR_INT */
1028 	PERF_BR_UNKNOWN,	/* X86_BR_IRET */
1029 	PERF_BR_COND,		/* X86_BR_JCC */
1030 	PERF_BR_UNCOND,		/* X86_BR_JMP */
1031 	PERF_BR_UNKNOWN,	/* X86_BR_IRQ */
1032 	PERF_BR_IND_CALL,	/* X86_BR_IND_CALL */
1033 	PERF_BR_UNKNOWN,	/* X86_BR_ABORT */
1034 	PERF_BR_UNKNOWN,	/* X86_BR_IN_TX */
1035 	PERF_BR_UNKNOWN,	/* X86_BR_NO_TX */
1036 	PERF_BR_CALL,		/* X86_BR_ZERO_CALL */
1037 	PERF_BR_UNKNOWN,	/* X86_BR_CALL_STACK */
1038 	PERF_BR_IND,		/* X86_BR_IND_JMP */
1039 };
1040 
1041 static int
1042 common_branch_type(int type)
1043 {
1044 	int i;
1045 
1046 	type >>= 2; /* skip X86_BR_USER and X86_BR_KERNEL */
1047 
1048 	if (type) {
1049 		i = __ffs(type);
1050 		if (i < X86_BR_TYPE_MAP_MAX)
1051 			return branch_map[i];
1052 	}
1053 
1054 	return PERF_BR_UNKNOWN;
1055 }
1056 
1057 /*
1058  * implement actual branch filter based on user demand.
1059  * Hardware may not exactly satisfy that request, thus
1060  * we need to inspect opcodes. Mismatched branches are
1061  * discarded. Therefore, the number of branches returned
1062  * in PERF_SAMPLE_BRANCH_STACK sample may vary.
1063  */
1064 static void
1065 intel_pmu_lbr_filter(struct cpu_hw_events *cpuc)
1066 {
1067 	u64 from, to;
1068 	int br_sel = cpuc->br_sel;
1069 	int i, j, type;
1070 	bool compress = false;
1071 
1072 	/* if sampling all branches, then nothing to filter */
1073 	if (((br_sel & X86_BR_ALL) == X86_BR_ALL) &&
1074 	    ((br_sel & X86_BR_TYPE_SAVE) != X86_BR_TYPE_SAVE))
1075 		return;
1076 
1077 	for (i = 0; i < cpuc->lbr_stack.nr; i++) {
1078 
1079 		from = cpuc->lbr_entries[i].from;
1080 		to = cpuc->lbr_entries[i].to;
1081 
1082 		type = branch_type(from, to, cpuc->lbr_entries[i].abort);
1083 		if (type != X86_BR_NONE && (br_sel & X86_BR_ANYTX)) {
1084 			if (cpuc->lbr_entries[i].in_tx)
1085 				type |= X86_BR_IN_TX;
1086 			else
1087 				type |= X86_BR_NO_TX;
1088 		}
1089 
1090 		/* if type does not correspond, then discard */
1091 		if (type == X86_BR_NONE || (br_sel & type) != type) {
1092 			cpuc->lbr_entries[i].from = 0;
1093 			compress = true;
1094 		}
1095 
1096 		if ((br_sel & X86_BR_TYPE_SAVE) == X86_BR_TYPE_SAVE)
1097 			cpuc->lbr_entries[i].type = common_branch_type(type);
1098 	}
1099 
1100 	if (!compress)
1101 		return;
1102 
1103 	/* remove all entries with from=0 */
1104 	for (i = 0; i < cpuc->lbr_stack.nr; ) {
1105 		if (!cpuc->lbr_entries[i].from) {
1106 			j = i;
1107 			while (++j < cpuc->lbr_stack.nr)
1108 				cpuc->lbr_entries[j-1] = cpuc->lbr_entries[j];
1109 			cpuc->lbr_stack.nr--;
1110 			if (!cpuc->lbr_entries[i].from)
1111 				continue;
1112 		}
1113 		i++;
1114 	}
1115 }
1116 
1117 void intel_pmu_store_pebs_lbrs(struct pebs_lbr *lbr)
1118 {
1119 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1120 	int i;
1121 
1122 	cpuc->lbr_stack.nr = x86_pmu.lbr_nr;
1123 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
1124 		u64 info = lbr->lbr[i].info;
1125 		struct perf_branch_entry *e = &cpuc->lbr_entries[i];
1126 
1127 		e->from		= lbr->lbr[i].from;
1128 		e->to		= lbr->lbr[i].to;
1129 		e->mispred	= !!(info & LBR_INFO_MISPRED);
1130 		e->predicted	= !(info & LBR_INFO_MISPRED);
1131 		e->in_tx	= !!(info & LBR_INFO_IN_TX);
1132 		e->abort	= !!(info & LBR_INFO_ABORT);
1133 		e->cycles	= info & LBR_INFO_CYCLES;
1134 		e->reserved	= 0;
1135 	}
1136 	intel_pmu_lbr_filter(cpuc);
1137 }
1138 
1139 /*
1140  * Map interface branch filters onto LBR filters
1141  */
1142 static const int nhm_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1143 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= LBR_ANY,
1144 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= LBR_USER,
1145 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= LBR_KERNEL,
1146 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
1147 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= LBR_RETURN | LBR_REL_JMP
1148 						| LBR_IND_JMP | LBR_FAR,
1149 	/*
1150 	 * NHM/WSM erratum: must include REL_JMP+IND_JMP to get CALL branches
1151 	 */
1152 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT] =
1153 	 LBR_REL_CALL | LBR_IND_CALL | LBR_REL_JMP | LBR_IND_JMP | LBR_FAR,
1154 	/*
1155 	 * NHM/WSM erratum: must include IND_JMP to capture IND_CALL
1156 	 */
1157 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT] = LBR_IND_CALL | LBR_IND_JMP,
1158 	[PERF_SAMPLE_BRANCH_COND_SHIFT]     = LBR_JCC,
1159 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT] = LBR_IND_JMP,
1160 };
1161 
1162 static const int snb_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1163 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= LBR_ANY,
1164 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= LBR_USER,
1165 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= LBR_KERNEL,
1166 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
1167 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= LBR_RETURN | LBR_FAR,
1168 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT]	= LBR_REL_CALL | LBR_IND_CALL
1169 						| LBR_FAR,
1170 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT]	= LBR_IND_CALL,
1171 	[PERF_SAMPLE_BRANCH_COND_SHIFT]		= LBR_JCC,
1172 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT]	= LBR_IND_JMP,
1173 	[PERF_SAMPLE_BRANCH_CALL_SHIFT]		= LBR_REL_CALL,
1174 };
1175 
1176 static const int hsw_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1177 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= LBR_ANY,
1178 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= LBR_USER,
1179 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= LBR_KERNEL,
1180 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
1181 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= LBR_RETURN | LBR_FAR,
1182 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT]	= LBR_REL_CALL | LBR_IND_CALL
1183 						| LBR_FAR,
1184 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT]	= LBR_IND_CALL,
1185 	[PERF_SAMPLE_BRANCH_COND_SHIFT]		= LBR_JCC,
1186 	[PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT]	= LBR_REL_CALL | LBR_IND_CALL
1187 						| LBR_RETURN | LBR_CALL_STACK,
1188 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT]	= LBR_IND_JMP,
1189 	[PERF_SAMPLE_BRANCH_CALL_SHIFT]		= LBR_REL_CALL,
1190 };
1191 
1192 /* core */
1193 void __init intel_pmu_lbr_init_core(void)
1194 {
1195 	x86_pmu.lbr_nr     = 4;
1196 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1197 	x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
1198 	x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
1199 
1200 	/*
1201 	 * SW branch filter usage:
1202 	 * - compensate for lack of HW filter
1203 	 */
1204 }
1205 
1206 /* nehalem/westmere */
1207 void __init intel_pmu_lbr_init_nhm(void)
1208 {
1209 	x86_pmu.lbr_nr     = 16;
1210 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1211 	x86_pmu.lbr_from   = MSR_LBR_NHM_FROM;
1212 	x86_pmu.lbr_to     = MSR_LBR_NHM_TO;
1213 
1214 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1215 	x86_pmu.lbr_sel_map  = nhm_lbr_sel_map;
1216 
1217 	/*
1218 	 * SW branch filter usage:
1219 	 * - workaround LBR_SEL errata (see above)
1220 	 * - support syscall, sysret capture.
1221 	 *   That requires LBR_FAR but that means far
1222 	 *   jmp need to be filtered out
1223 	 */
1224 }
1225 
1226 /* sandy bridge */
1227 void __init intel_pmu_lbr_init_snb(void)
1228 {
1229 	x86_pmu.lbr_nr	 = 16;
1230 	x86_pmu.lbr_tos	 = MSR_LBR_TOS;
1231 	x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
1232 	x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
1233 
1234 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1235 	x86_pmu.lbr_sel_map  = snb_lbr_sel_map;
1236 
1237 	/*
1238 	 * SW branch filter usage:
1239 	 * - support syscall, sysret capture.
1240 	 *   That requires LBR_FAR but that means far
1241 	 *   jmp need to be filtered out
1242 	 */
1243 }
1244 
1245 /* haswell */
1246 void intel_pmu_lbr_init_hsw(void)
1247 {
1248 	x86_pmu.lbr_nr	 = 16;
1249 	x86_pmu.lbr_tos	 = MSR_LBR_TOS;
1250 	x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
1251 	x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
1252 
1253 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1254 	x86_pmu.lbr_sel_map  = hsw_lbr_sel_map;
1255 
1256 	if (lbr_from_signext_quirk_needed())
1257 		static_branch_enable(&lbr_from_quirk_key);
1258 }
1259 
1260 /* skylake */
1261 __init void intel_pmu_lbr_init_skl(void)
1262 {
1263 	x86_pmu.lbr_nr	 = 32;
1264 	x86_pmu.lbr_tos	 = MSR_LBR_TOS;
1265 	x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
1266 	x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
1267 
1268 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1269 	x86_pmu.lbr_sel_map  = hsw_lbr_sel_map;
1270 
1271 	/*
1272 	 * SW branch filter usage:
1273 	 * - support syscall, sysret capture.
1274 	 *   That requires LBR_FAR but that means far
1275 	 *   jmp need to be filtered out
1276 	 */
1277 }
1278 
1279 /* atom */
1280 void __init intel_pmu_lbr_init_atom(void)
1281 {
1282 	/*
1283 	 * only models starting at stepping 10 seems
1284 	 * to have an operational LBR which can freeze
1285 	 * on PMU interrupt
1286 	 */
1287 	if (boot_cpu_data.x86_model == 28
1288 	    && boot_cpu_data.x86_stepping < 10) {
1289 		pr_cont("LBR disabled due to erratum");
1290 		return;
1291 	}
1292 
1293 	x86_pmu.lbr_nr	   = 8;
1294 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1295 	x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
1296 	x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
1297 
1298 	/*
1299 	 * SW branch filter usage:
1300 	 * - compensate for lack of HW filter
1301 	 */
1302 }
1303 
1304 /* slm */
1305 void __init intel_pmu_lbr_init_slm(void)
1306 {
1307 	x86_pmu.lbr_nr	   = 8;
1308 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1309 	x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
1310 	x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
1311 
1312 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1313 	x86_pmu.lbr_sel_map  = nhm_lbr_sel_map;
1314 
1315 	/*
1316 	 * SW branch filter usage:
1317 	 * - compensate for lack of HW filter
1318 	 */
1319 	pr_cont("8-deep LBR, ");
1320 }
1321 
1322 /* Knights Landing */
1323 void intel_pmu_lbr_init_knl(void)
1324 {
1325 	x86_pmu.lbr_nr	   = 8;
1326 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1327 	x86_pmu.lbr_from   = MSR_LBR_NHM_FROM;
1328 	x86_pmu.lbr_to     = MSR_LBR_NHM_TO;
1329 
1330 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1331 	x86_pmu.lbr_sel_map  = snb_lbr_sel_map;
1332 
1333 	/* Knights Landing does have MISPREDICT bit */
1334 	if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_LIP)
1335 		x86_pmu.intel_cap.lbr_format = LBR_FORMAT_EIP_FLAGS;
1336 }
1337